NVIDIA introduces a synthetic-video detector, not an automated verdict

NVIDIA presented the Synthetic Video Detector NIM at SIGGRAPH on July 20, 2026. NVIDIA’s official AI for Media update still lists it as a private-access product, while the documentation provides a hosted API demonstration. It is not a generally available consumer verification tool and should not be described as a final authority on authenticity.

The microservice analyses video frame by frame and returns synthetic-content probability scores, plus an aggregate probability for the complete clip. NVIDIA says it looks for statistical and frequency-domain traces associated with generative pipelines. In NVIDIA’s own testing, reported accuracy reached 92% for uncompressed video, 87% at 15% compression and 82% at 50% compression. The company also reports 1080p processing in as little as 22 ms on RTX systems and roughly 30 ms on L40 GPUs. These are vendor-reported results, without independent validation or a complete error matrix across every content category.

The practical value is triage: a newsroom can prioritise clips for deeper review without automatically sending all source material to an outside service. On-premises, edge and approved air-gapped deployment can also matter for source confidentiality. The score remains only one signal. A false positive can cast doubt on authentic footage, while a false negative can allow synthetic video through. Provenance, source checks, metadata, frame analysis and journalistic confirmation remain essential.

We are watching for a clear move from private access to commercial availability, documentation of the evaluation sets, false-positive and false-negative rates, behaviour on legitimately edited footage, and performance against generative models released after the detector was trained. We will treat NVIDIA’s figures as vendor claims until independently tested.